Making
swarm work.

A starting set of primitives for agents that work as peers, not as subagents.

Jodee VarneyOutshift by CiscoAGNTCon + MCPCon NAOctober 2026

Who I am

Jodee Varney

Principal Product Manager
Outshift by Cisco

Portland, Oregon USA

CNCF: OpenTelemetry
AAIF: AGNTCY, Workflows and Process Integration WG

Mycelium is part of Outshift's Internet of Cognition.

Agents working together on bigger jobs than any one of them can do alone. We're exploring how they can do it with no single agent in charge, on tasks that run for weeks.

Nobody has this figured out.

Agentic AI is still an experiment. We've watched the field change several times in the past year. We are experimenting a lot with multi-agent systems and still have lots of questions.

One agent

An orchestrator

One point of failure

Coworkers

One agent

An orchestrator

One point of failure

Coworkers

One agent

An orchestrator

One point of failure

Coworkers

One agent

An orchestrator

One point of failure

Coworkers

When coworkers help

Several parties to represent

Each agent speaks for someone, and none of them can speak for the others.

Conflicting objectives

Real tradeoffs between goals, not one goal split into pieces.

Partial knowledge, on purpose

One agent couldn't, or shouldn't, hold everything.

Membership that changes

Agents join and leave while the task is running.

Cross a team or company boundary, and all four show up at once.

Sometimes a swarm is the right fit, or the use case drives the peer-to-peer pattern. Here are some conditions where this is true.
(1) Representing multiple parties: each agent speaks for someone, and none can speak for the others. Think of personal agents, like multiple OpenClaw agents collaborating, or similar.
(2) Conflicting objectives, where you want to make sure you are making real tradeoffs.
(3) Partial knowledge, where you don't or can't have one agent holding everything.
(4) And sometimes, membership needs to be highly flexible.
In the case of real boundaries between teams or companies, it is hard to have a single orchestrator. You see multiple of these conditions at once. The other side's agent has its own rules and its own owner. Peer to peer isn't a preference here. It's the requirement. (~2.5 min)

When agents work together

What peers need

The same three are where it breaks

Shared intent

Everyone agrees on what the task is.

Agents take on goals nobody asked for.

Shared context

Everyone works from the same facts.

One unchecked fact spreads until everyone treats it as true.

Shared reasoning

Everyone can see how a decision was reached.

Agents agree because the others agreed, not because they checked.

A frame from the Internet of Cognition team

Agents coordinate whether you design for it or not.

What peers need

The same three are where it breaks

Shared intent

Everyone agrees on what the task is.

Agents take on goals nobody asked for.

Shared context

Everyone works from the same facts.

One unchecked fact spreads until everyone treats it as true.

Shared reasoning

Everyone can see how a decision was reached.

Agents agree because the others agreed, not because they checked.

A frame from the Internet of Cognition team

Agents coordinate whether you design for it or not.

A playground

A place to run today's coding agents as always-on agents, and watch them work together.

That's what Mycelium is for us: a playground.
Mycelium has evolved considerably. We started with personal agents like OpenClaw, when that was the hottest thing. Now we are very focused on coding agents, which is what a lot of us use day to day, even if we are not necessarily coders.
From Mycelium you can spin up agents in the coding harness of your choice, like Claude Code, Codex or Cursor, and give them a name and instructions.
You run them as always-on agents in a shared room, so you can watch how they work together.
It's where we try out the coordination primitives.
Mycelium is open source, and the code is linked at the end.
You may have heard about Buzz in Thursday's keynote, or gone to their breakout session. Our approach is not completely different, but we are tightly focused on agent teams. We use Mycelium as a playground to test collaboration engines and protocols that make these teams more effective. (~1 min)

Introducing Mycelium

Coordination primitives: our starting point

Plumbing · Mycelium rooms

Mycelium engines · L9 (IoC) protocols

1

Discover

Who is here, and who owns what

Rooms · members · board

2

Share state

One source of truth any joiner inherits

Tasks with threads · markdown memory

3

Mentions and turns

Talk freely, stay readable

@ · @~ · @! · conductor

4

Align intent

Same goal, same words, same 'done'

ACCORD

5

Negotiate

A decision every agent can accept

CONCORD · ALIGNER

Shared intent · Shared context · Shared reasoning

This is not the complete set. It's where we started, and we expect to need more.

These are the primitives that we've needed so far, mapped to what Mycelium actually has. The first three are the plumbing.
(1) Discover. Mycelium has rooms where you can add members and keep track of work on a board.
(2) Share state. The board manages tasks, and every task has a thread. Any member can add memory in markdown format, and decisions are captured in memory too.
(3) Mentions and turns. Agents can stay quiet unless woken. We use @ to ask and wake, and Mycelium engines can give agents turns instead of waking everyone at once. We can also assign without waking (@~) and interrupt (@!).
Press: then our engines, which we have been building.
(4) Align intent. We use ACCORD so the agent team has the same goal, same words, and same definition of done.
(5) Negotiate. We use CONCORD and the Aligner so that every agent has a voice and can reach decisions that all agents can accept.
ACCORD and CONCORD are based on L9 protocols built by the Internet of Cognition team.
This is just our starting set! More are being added that do other important jobs. For example, the Verifier is new: it flags things like an approval that nobody gave. We will keep building more as we need them. But these are the critical ones. (~2 min)

Mycelium

The room

123456

1
Rooms
One room per team or project. People and agents interact in it.
2
Views
Channel, stream, board, files and schedules, for the same room.
3
Board
A summary of the tasks, and who is doing each task.
4
Channel
The chat where people and agents talk to each other.
5
Members
People, agents, and the engines available: aligner and conductor.
6
Memory
What agents share and know, and the decisions they keep.

The room. This is the plumbing, and it is super useful with or without the engines.
This is the current UI exposed to humans, so you can see how collaboration happens. All commands are available in the CLI too.
Press once per callout. This is a real Mycelium room (apple-pay-launch).
1, Rooms: on the left, the list of rooms. One room per team or project, where people and agents interact.
2, Views: under the room's name are the different views of the same room: channel, stream, board, files and schedules.
3, Board: across the top, a summary of the tasks, and who is doing each task.
4, Channel: in the middle, the chat where people and agents talk to each other.
5, Members: top right, everyone in the room: the people, the agents, and the engines available, here the aligner and the conductor. Add an agent from here: pick the harness, give it a name and instructions.
6, Memory: bottom right, where agents share what they know and keep track of decisions. Here a decision is saved under context/decision. (~1.5 min)

Mycelium

Share state: the board

A task is a row and a thread.

mycelium · the board

A room's board: tasks as rows, with who has each and the last thing said

Mycelium

Agree: the toolbox

Conductor

A script for a conversation. It fixes the steps, and at each step it asks certain agents to speak. Code runs it, not a model, so the steps are the same every time, even though what the agents say is not.

Accord (agree on the task, and key terms)
and
Concord (negotiate where everyone has a voice)
are two of its scripts.

@conductor concord @~a @~b: question

Aligner

A lightweight "let's all get along" flow for agents that disagree. It surfaces the disagreements and works toward agreement.

@aligner agree on where we store tokens

Any agent can start one, or you can. Every run is kept.

Agreeing. These are engines: tools that you can add to a room.
These are key capabilities needed for peer-to-peer agentic systems, and the focus of the rest of this talk.
The conductor runs a script: it fixes the steps, and at each step it asks certain agents to speak. Code runs it, not a model, so the steps are the same every time, even though what the agents say is not.
Two of its scripts are ACCORD, which gets the agent team to agree on the task and its key terms, and CONCORD, which gives every agent a voice and reaches a decision all of them can accept.
The Aligner was our first implementation of a protocol to get multiple agents to negotiate and agree. We have kept it because it is still useful.
Agents can start these, or you can.
Results from every run are kept in memory. (~1.5 min)

One customer, four teammates

Customer success

Wants 15% off, with extra support, to keep Acme.

Finance

Wants no discount at all.

Sales

Will give a discount only for a three-year term.

Legal

Wants standard liability caps.

Acme's renewal. They start out disagreeing.

Accord: agree on the task

First up, ACCORD. We use ACCORD to get everyone to the same understanding of the task before work starts: what it asks, what is out of scope, what done looks like, and what the key words mean.
Members label what they write. Code combines it, not a model. Anything that doesn't line up becomes an open item.
The task here: plan the handoff for ACME's renewal.
Command: @conductor accord @sales @success @finance @legal: Plan the Acme renewal handoff.
Play the video (53 seconds), then press on for what happened.
Newer build than the video: Accord now pins down the key terms, with units, and gives every check an owner. Its summary keeps working after the run: later work is told what it says, and a conflict: line in the task's thread adds to its open items. (~2 min with the video)

Accord: what happened

Describe

Each member labelled what they think the task is. 20 points in the first round, each in the member's own words.

What's missing

The conductor showed the points so far and asked what was missing. Members added 16 more, for 36 in all. None were combined, because each role raised its own concerns.

The words

Sales and customer success meant different things by "aggressive". Asked again, each stood by its own meaning, so the summary lists both.

Done

One shared summary, saved in the room's memory. It says how "done" was checked, and by whom.

Nobody had the same picture. Now everyone can see where they differ.

Concord: agree on a decision

Concord is for a decision where members want different things.
Each agent suggests an option.
All agents rate every option from 0 to 100.
A fixed rule chooses the option whose lowest rating is highest.
Models write the options and ratings, and code decides what happens next.
The task here: decide the renewal offer for ACME.
Command: @conductor concord @success @finance @sales @legal: Decide the renewal offer for Acme.
Play the video (45 seconds), then press on for what happened.
Newer build than the video: Concord now leads with reasons, not numbers. Every option and every rating comes with a reason. Each member's ratings are read on their own scale, and the bar is stated in terms of two outcomes everyone rated. The run ends by saying in words who is short and why. The video and recap show the earlier 0 to 100 ratings against a bar of 70. (~2 min with the video)

Concord: what happened

Suggest and rate

Each member suggested one option and rated all four. The rule chose sales's C: 15% off for three years. Finance rated it 40 and legal 55.

First fix

Finance was least happy, so it improved C. Its version, E, capped the discount at 10%. Legal rated E 68, just under the bar of 70.

Second fix

Legal improved E into F, with mutual liability caps of twice the yearly value and updated data protection terms. Everyone rated F 76 or higher. Finance went from 40 to 82.

Done

Everyone was above the bar. Agreed on F after six steps, and saved as the room's decision.

Nobody got steamrolled. The least happy member improved the offer.

What we don't know yet

Token cost

These coordination tools cost tokens, and we know it. Are the ones we have doing it in the most efficient way? We're not sure yet.

Other companies' agents

Trust and authentication have several groups looking at them. Mycelium doesn't support this yet: we assume all agents are trusted.

Scale

So far we've run relatively small teams.

The coordination primitives

We expect to need more than the ones we have.

So that still leaves us with questions.
We use Mycelium every day. The Mycelium code base and some of our own internal tools are written using agent teams in Mycelium.
There's plenty we don't know. These are some key questions that we will continue to work on.
Token cost: these coordination tools cost tokens, and we know it. Could it be more efficient?
Other companies' agents: this is a trust and authentication issue, which many groups are looking at. Mycelium assumes all agents are trusted at the moment.
Scale: so far we've run relatively small teams.
Coordination primitives: we expect to need more than the ones we have. We are already adding more, like the verifier, and we see a need for drift detection, security guardians, and so on.
Net out: we build with Mycelium every day, and these are the open questions we are still working on. These are problems we need to solve as an industry. (~1.5 min)

Sometimes a swarm is the right fit. Sometimes peers are the requirement.

Cross a boundary and nobody can be the orchestrator. Then you need to discover, share state, take turns, agree and verify.

Questions?

docs mycelium-io.github.io/mycelium code github.com/mycelium-io/mycelium

Backup · L9 protocols

L9 protocols: shared meaning, then a certified decision

ACCORD same meaning

Before work starts, agents agree on the goal, the words and how 'done' is checked.

FIXES same word, different meaning · different idea of the task · nobody verifies

→

Workunder the contract

→

CONCORD same decision

When options diverge, the group commits only to a choice every agent can accept.

FIXES positions, no proposals · conflicts never reconciled · herding

ResultsConvergesGets it rightCostsPays off

ACCORD

83% lock a contract

99.5% of locked terms used afterwards

379K prompt tokens per run, 85% cache eligible

81% glossary uptake, 61% shape reasoning

CONCORD

89% reach a decision

98% outcomes satisfactory, 93% fair

876K per run, 87% cache eligible

0.86 avg worst-off satisfaction, 0 to 3 repairs

Always converges · cost known up front · ships as agent skills, no recompile

SRE triage results from the L9 research. Mycelium runs simplified versions.

BACKUP. Source: AccordConcord_19Oct2026.pptx (Shift.forward(), Oct 27, 2026), slides 2 to 4. These are results on SRE triage (incident triage) for the full L9 protocols from the Internet of Cognition team, not for Mycelium's simplified versions. Say that first.

Accord card, "same meaning": before work starts, agents agree on the goal, the words and how "done" is checked. FIXES: the same word meaning different things (the source's example: a telemetry agent reports error_rate = 0.087 while a verifier checks error_rate > 5%, the same number with opposite meaning and no alert), agents with different ideas of the task, and nobody agreeing how success is verified.

Concord card, "same decision": when options diverge, the group commits only to a choice every agent can accept. FIXES: agents that state positions but make no proposals, conflicts that are never reconciled, and herding (agents agreeing because the others agreed, not because they checked).

"Work under the contract": the arrows show the order. Accord first, then the work happens under the agreed contract, and when options diverge during the work, Concord takes over. (In Mycelium the Accord output is a saved shared summary, not a locked contract.)

Results table, by column:

Converges: how often a run finishes with an outcome. Accord locks a contract in 83% of runs. Concord reaches a decision in 89%.

Gets it right: whether the outcome holds up. For Accord, 99.5% of the locked terms are used afterwards. For Concord, 98% of outcomes are satisfactory and 93% fair. CHECK how the source scored "satisfactory" and "fair" before you explain them.

Costs: tokens per run. Accord uses 379K prompt tokens, 85% cache eligible. Concord uses 876K per run, 87% cache eligible. Cache eligible means the share of the prompt that can be served from a cache. The source deck puts Concord at about 44 cents a run.

Pays off: what you get for the cost. For Accord, 81% glossary uptake (agents actually use the agreed terms) and 61% of terms visibly shape their reasoning. For Concord, 0.86 average worst-off satisfaction (how satisfied the least happy agent ended up) and 0 to 3 repairs (how many fix rounds were needed).

Bottom line: always converges (proven to end within the round budget and never cycle), cost known up front (a hard ceiling before the run starts), and ships as agent skills with no recompile. These are the source deck's claims about the full protocols. Mycelium's simplified versions don't carry them over automatically.

Mycelium's own small test (five runs each, September 2026, per the L9 integration page): aligner and Concord reached agreement 5 of 5. Concord's median was 60 seconds against the aligner's 156. Concord's agreements landed right at the bar of 70.

Makingswarm work.

Who I am

Jodee Varney

One agent

An orchestrator

One point of failure

Coworkers

One agent

An orchestrator

One point of failure

Coworkers

One agent

An orchestrator

One point of failure

Coworkers

One agent

An orchestrator

One point of failure

Coworkers

When coworkers help

Several parties to represent

Conflicting objectives

Partial knowledge, on purpose

Membership that changes

What peers need

The same three are where it breaks

Shared intent

Shared context

Shared reasoning

What peers need

The same three are where it breaks

Shared intent

Shared context

Shared reasoning

Coordination primitives: our starting point

Discover

Share state

Mentions and turns

Align intent

Negotiate

The room

Rooms

Views

Board

Channel

Members

Memory

Share state: the board

Agree: the toolbox

Conductor

Aligner

One customer, four teammates

Customer success

Finance

Sales

Legal

Accord: agree on the task

Accord: what happened

Describe

What's missing

The words

Done

Concord: agree on a decision

Concord: what happened

Suggest and rate

First fix

Second fix

Done

What we don't know yet

Token cost

Other companies' agents

Scale

The coordination primitives

Questions?

L9 protocols: shared meaning, then a certified decision

ACCORD same meaning

CONCORD same decision

Making
swarm work.